This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: RFC: eliminate some dead stores


Dan Nicolaescu writes:
> Stores to structs that are passed by value, or to struct local vars
> are dead, but they are not eliminated currently. 

Not just those passed by value: in many cases, temporary objects passed
by const reference to inline functions generate dead stores.  For some
reason, we usually wind up with more dead stores on Sparc than on ix86.
Not sure why.

I believe this to be the #1 performance problem in GNU C++.  Any STL-like
code that uses iterators with more than one data member generates tons
of dead stores.  As I've said before, the Stepanov benchmark is misleading
as it only tests the cases we're good at.

[ patch deleted ]
I'm not qualified to comment on your patch, knowing little about flow.c,
your idea seems reasonable but a bit hackish.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]